RT info:eu-repo/semantics/article
T1 A predictive hybrid reduced order model based on proper orthogonal decomposition combined with deep learning architectures
A1 Abadía Heredia, R.
A1 López Martín, Manuel
A1 Carro Martínez, Belén
A1 Arribas Sánchez, Juan Ignacio
A1 Pérez, José Miguel
A1 Le Clainche, Soledad
K1 Reduced order models
K1 Deep learning architectures
K1 POD
K1 Modal decompositions
K1 Neural networks
K1 Fluid dynamics
K1 33 Ciencias Tecnológicas
K1 12 Matemáticas
AB Solving computational fluid dynamics problems requires using large computational resources. The computa-tional time and memory requirements to solve realistic problems vary from a few hours to several weeks withseveral processors working in parallel. Motivated by the need of reducing such large amount of resources(improving the industrial applications in which fluid dynamics plays a key role), this article introduces a newpredictive Reduced Order Model (ROM) applied to solve fluid dynamics problems. The model is based onphysical principles and combines modal decompositions with deep learning architectures. The hybrid ROM,reduces the dimensionality of a database via proper orthogonal decomposition (POD), extracting the dominantfeatures leading the flow dynamics of the problem studied. The number of degrees of freedom are reducedfrom hundred thousands spatial points describing the database to a few (20–100) POD modes. Firstly, PODdivides the spatio-temporal data into spatial modes and temporal coefficients (or temporal modes). Next, thetemporal coefficients are integrated in time using convolutional or recurrent neural networks. The temporalevolution of the flow is approximated after combining the spatial modes with the new temporal coefficientscomputed. The model is tested in two complex problems of fluid dynamics, the three-dimensional wake of acircular cylinder and a synthetic jet. The hybrid ROM uses data from the initial transient stage of numericalsimulations to predict the temporally converged solution of the flow with high accuracy. The speed-up factorcomparing the time necessary to obtain the predicted solution using the hybrid ROM and the numerical solveris ∼140–348 in the synthetic jet and ∼2897–3818 in the three dimensional cylinder wake. The robustness shownin the results presented and the data-driven nature of this ROM, make it possible to extend its application toother fields (i.e. video and language processing, robotics, finances)
PB Elsevier
SN 0957-4174
YR 2022
FD 2022
LK https://uvadoc.uva.es/handle/10324/54123
UL https://uvadoc.uva.es/handle/10324/54123
LA eng
NO Expert Systems with Applications, 2022, vol. 187, p. 115910
NO Producción Científica
DS UVaDOC
RD 09-may-2024